Embeddings, the foundation of retrieval.

using LMKit.Model;
using LMKit.Embeddings;

var model    = LM.LoadFromModelID("embeddinggemma-2");
var embedder = new Embedder(model);

float[] vec = embedder.GetEmbeddings("Machine learning is fascinating.");
Console.WriteLine($"dimensions: {vec.Length} (matches embedder.EmbeddingSize = {embedder.EmbeddingSize})");

var texts = new[]
{
    "The cat sat on the mat.",
    "A feline rested on the rug.",
    "Stock market closed higher.",
};

// One call, three vectors. GPU batches them efficiently.
float[][] vecs = embedder.GetEmbeddings(texts);

// 0 and 1 are semantically close; 2 is far away.
float simCatRug   = VectorOperations.CosineSimilarity(vecs[0], vecs[1]);
float simCatStock = VectorOperations.CosineSimilarity(vecs[0], vecs[2]);

Console.WriteLine($"cat ~ rug:   {simCatRug:F3}");
Console.WriteLine($"cat ~ stock: {simCatStock:F3}");

using LMKit.Media.Image;

// One model for text and images.
var embedder = new Embedder(LM.LoadFromModelID("embeddinggemma-2"));

// Index images.
var index = new List<(string Path, float[] Vector)>();
foreach (var path in Directory.EnumerateFiles("./photos", "*.jpg"))
{
    using var image = ImageBuffer.LoadAsRGB(path);
    index.Add((path, embedder.GetEmbeddings(image)));
}

// Query with TEXT. Retrieves IMAGE results.
float[] query = embedder.GetEmbeddings("sunset over the ocean");
var hits = index
    .OrderByDescending(e => Embedder.GetCosineSimilarity(query, e.Vector))
    .Take(5);